Professional Documents
Culture Documents
www.elsevier.com/locate/jprocont
Abstract
Even though employed widely in industrial practice, the popular PID controller has weaknesses that limit its achievable performance,
and an intrinsic structure that makes tuning not only more complex than necessary, but also less transparent with respect to the key
attributes of the overall controller performance, namely: robustness, set-point tracking, and disturbance rejection. In this paper, we propose an alternative control scheme that combines the simplicity of the PID controller with the versatility of model predictive control
(MPC) while avoiding the tuning problems associated with both. The tuning parameters of the proposed control scheme are related
directly to the controller performance attributes; they are normalized to lie between 0 and 1; and they arise naturally from the formulation in a manner that makes it possible to tune the controller directly for each performance attribute independently. The result is a
controller that can be designed and implemented much more directly and transparently, and one that outperforms the classical PID controller both in set-point tracking and disturbance rejection while using precisely the same process reaction curve information required to
tune PID controllers. The design, implementation and performance of the controller are demonstrated via simulation on a nonlinear
polymerization process.
2005 Elsevier Ltd. All rights reserved.
Keywords: Regulatory control; SISO controller design; PID control; Controller tuning; Model predictive control
1. Introduction
The PID controller is by far the most commonly used
controller in industrial practice. Based on a survey of the
status of industrial process controllers, Desborough and
Miller [1] estimated that 98% of the controllers in a median chemical plant are PID controllers. Furthermore, it is
estimated that there are perhaps only 510% of control
loops for which a typical practitioner will not consider
applying the basic single input, single output (SISO) PID
control strategy [2]. The popularity of the PID controller
is normally attributed to its simple structure and ease of
implementation in hardware and software. However, the
PID controller structureeven the digital implementationis based on century-old concepts that were dictated
*
Corresponding author. Tel.: +1 302 831 4504; fax: +1 302 831 1048.
E-mail address: ogunnaike@che.udel.edu (B.A. Ogunnaike).
0959-1524/$ - see front matter 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2005.08.001
500
Keas
us
ss 1
k 0; 1; 2; . . .
where
Dt
a e s
Dt
b K 1 e s
a
m round
Dt
where
1
lk; m
m
X
i1
ai uk i
represents the m-term weighted sum of previous control actions taken during the delay period m. Note that Eq. (4)
now explicitly contains u(k), the current control action. If
we now require that from the current time step onwards,
and over the next N-step horizon, only one control move,
u(k), is allowed, i.e.,
uk i uk;
i 1; 2; . . . ; N
16i6N
where
501
error therefore could be due to any of the following components: parametric uncertainty arising from wrongly estimating a0, b0, m0 as a, b, m respectively; structural uncertainty
arising because the dynamics represented by dS(k) is absent
from our simplistic model (2); unmodelled disturbances because dD(k), assumed to be unmeasurable, is excluded from
the model; and random measurement noise.
By subtracting Eq. (2) from (10) and recalling the denition of the model error in (9), we obtain:
ek 1 aek Dayk Dbuk m bDuk m
dS k dD k nk
11
where
Da a0 a
Db b0 b
1 ai
gi
1a
Duk m uk m0 uk m
ek 1 aek dP k dS k dD k nk
aek dB k dR k
The approach by which the basic model prediction obtained in Eq. (7) is updated using available process measurements is a key dierentiator of this control strategy
from other model predictive control schemes; it will therefore be described in some detail.
2.2.1. Modeling error characterization
A simplistic FOPDT model (or any other model for that
matter) cannot be expected to represent the true process
dynamics perfectly; but the typical modeling error obtained
from available plant data, y(k), as
ek yk ^y k
is a combination of several apparently inseparable components. Our approach requires rst, a careful characterization of the various components that contribute to the
observed modeling error, to provide the basis for a judicious estimation of what can realistically be estimated.
We consider the nominal model given in Eq. (2) to be an
approximation of the unknown reality given by
yk 1 a0 yk b0 uk m0 dS k dD k nk
10
where a0, b0 and m0 represent the true, but unknown,
parameters associated with the rst-order portion of the
true process dynamics; dS represents the higher order, possibly nonlinear, process dynamics; dD represents the
unmodelled external disturbances; and n is the random
measurement noise component. The observed modeling
12
where all the bias-inducing components have been consolidated into dB(k), leaving dR(k) as the non-biasing residual.
2.2.2. Model error resolution
We now resolve the model error e(k) into two components, em(k) and eD(k), according to
ek em k eD k
15
16
17
502
This leads to an update strategy in which eD(k) is estimated as ^eD k from the available model error e(k) such
that the expected value of the complementary component,
em(k), is required to be exactly zero. From here, the model
prediction update strategy consists of two parts:
(1) Current disturbance eect estimation: obtaining from
e(k), an estimate of eD(k), the current disturbance
eect as ^eD k, and
(2) Future disturbance eect predictions: using ^eD k to
obtain ^eD k j j k, the prediction of the future values of the disturbance eect over the extended horizon j = 1, 2, . . . , N.
2.2.3. Current disturbance eect estimation
From the preceding discussion, we state the disturbance
eect estimation problem formally as follows:
Given e(k), estimate eD(k) such that em(k) has a mean
value of zero.
Probabilistic descriptions: we employ the following probabilistic descriptions to characterize the uncertainty associated with the various components of this estimation
problem:
(1) The belief in the integrity of the nominal model is
quantied by a probability density function (pdf)
for em(k) having a zero mean (by denition) and
variance r2m . A reasonably objective estimate of r2m
can be obtained from the nominal modeling exercise.
In general, the greater the inherent model uncertainties, the higher the value of r2m ; the converse is also
true.
(2) The composite dynamic behavior of the bias-inducing
component (including unmodelled disturbances) is
considered to evolve according to the following stochastic process:
dB k dB k 1 wk
18
19
20
$eD(k) therefore exhibits a stationary AR(1) behavior (for jaj < 1) with an expected value of zero [20].
The behavior of the composite disturbance eect
may therefore be characterized by a probability distribution on $eD(k) having zero mean and a variance
m 2.
Bayesian estimation: it is now important to observe that
there are two ways by which eD(k) can be estimated:
(i) From the plant/model mismatch characteristics in
Eq. (15), as ^eD k Eek. This follows directly
from the denition of em(k) and the requirement that
E[em(k)] = 0.
(ii) From the characteristics of the unmodelled disturbances where, as established above, $eD(k), dened
as
reD k eD k eD k 1
21
22
23
24
NeD k 1; m
25
26
where
hR
r2m
r2m m2
27
32
In practice, the correlation structure of the dB(k) sequence may be more complicated than that of the convenient random walk process, with the net implication that
the operating equation for the evolution of the composite
disturbance eect, Eq. (20), should contain additional
dynamics. However, in the absence of further legitimate
information for objectively introducing structural modications to this equation, we retain the structure, but replace
the xed parameter a with a tuning parameter, (1 hD)
so that Eq. (20) now reads:
r^eD k 1 1 hD r^eD k wk
33
1 hD
j
1 1 hD r^eD k;
hD
m16j6mN
34
r^eD k 1 j k ar^eD k
28
503
35
and, in general:
r^eD k j j k aj r^eD k
29
Since
j
X
l1
30
504
36
16j61
37
Here hT is clearly the set-point tracking tuning parameter such that for hT 0, almost instantaneous response to
set-point changes is demandedan aggressive policy; conversely, hT 1 indicates a conservative set-point tracking
policy, since it requires a more sluggish approach to the
set-point. (Note that the extreme value of hT = 1 is an invalid and meaningless choice for which all set-point changes
are completely ignored.)
The control action u(k) may now be computed such that
the predicted process output deviation from the reference
trajectory is minimized over the N-step horizon beyond
the delay period, m. For a least-squares objective, the optimization problem is given by
N
X
2
min
y k i ~y k m i
38
uk
i1
By dening
ri k y k i ~y k m i;
39
40
or,
ri k wi k bgi uk
41
where
wi k y k i ami ^y k ai1 blk; m
^eD k m i j k
42
represents the conditional projected errorthe performance decit that will result if u(k) = 0, i.e., it is the predicted output deviation from desired trajectory, when the
contribution from the yet undetermined control action
u(k) is excluded. From Eq. (41), we may now see that the
optimization problem posed in (38) aims to determine the
value of u(k) that best makes up the indicated performance
decit. The analytical solution to the optimization problem
is the closed-form expression:
PN
1 i1 gi wi k
uk
43
PN 2
b
i1 gi
This is the analytic expression for computing the control
action; it contains only one nal tuning parameter yet to
be speciedN, the prediction horizon length beyond the
time delay.
N1Dt
s
44
s
ln1 hA
Dt
45
Thus, hA 0 will result in a small horizon length (aggressive controller) and hA 1 will yield a conservative controller employing a long prediction horizon.
Thus with the analytic expression in Eq. (43), the RTDA controller computes, at each time instant k, the control
action required to minimize, over an extended horizon of
length N (as determined by the tuning parameter hA), the
deviation of the predicted process output from the desired
trajectory (as determined by the tuning parameter hT), after
the output prediction has been updated to reect the eect
of plant/model mismatch (as determined by the parameters
hR to estimate the disturbance eect, and hD to predict its
evolution into the future), and conditioned on the fact that
only one control move is allowed over the entire horizon
length.
As with other model predictive control schemes, the
strategy is to carry out the computation indicated in Eq.
(43) at every time instant, in a receding horizon fashion
[13,14].
2.4. Control scheme summary and discussion
We now summarize the key aspects of the control
scheme presented above.
2.4.1. Requirements
(1) Process model
gs
Keas
ss 1
46
505
es
2s 1
51
506
Process Output
Process Output
1.2
1
0.8
0.6
0.4
T=D=A=0.3
=0.3
R
=0.6
R
R=0.9
0.2
0
Setpoint tracking
Robustness
10
(a)
15
0.8
0.6
0.4
R=D=A=0.3
T=0.1
T=0.9
T=0.95
0.2
0
20
(b)
Time
10
15
20
25
Time
1.4
1.2
Overall aggressiveness
Disturbance rejection
1.2
Process Output
Process Output
1
1
0.8
0.6
0.4
R= T= A=0.3
0.6
0.4
R=D= T=0.3
D=0.01
D=0.1
=0.9
0.2
0.8
0
0
(c)
10
15
20
A=0.1
A=0.5
=0.9
0.2
0
25
Time
10
(d)
15
20
25
Time
Fig. 1. Eect of tuning parameters on controller performance attributes. The arrows indicate the direction of increasing parameters.
15
NAMW.
Kg/Kmol (x103)
5
True process
FOPDT fit
0
0
0.5
1.5
Time. hour
2.5
0.5
1.5
Time. hour
2.5
0.02
3.2. Implementation
The procedure for implementing the RTD-A controller
is now illustrated via simulation on a process of industrial relevance. The process in question is the isothermal
polymerization reactor discussed in Maner et al. [28], where
initiator ow rate is used to control number average
molecular weight (NAMW). The nonlinear, four-state,
state-space model (available in the cited reference) is used
to represent the true plant. Since measurements of actual
NAMW are not available on-line, we consider instead the
more practical case where an on-line viscometer provides
a surrogate measurement that correlates well with NAMW,
but with a modest measurement delay of 0.1 h.
The approximate FOPDT model
10
0.015
0.01
0.005
0
0
NAMW. Kg/Kmol
4
3.5
3
2.5
x 10
0.5
1.5
Time. hour
0.02
0.015
0.01
0.005
0
0.5
(a)
1.5
Time. hour
2.5
2.6
2.4
2.2
2
2.5
Initiator flow rate. m3/hr
NAMW. Kg/Kmol
x 10
4.5
507
0.5
1.5
Time. hour
2.5
0.5
1.5
Time. hour
2.5
0.02
0.015
0.01
0.005
0
0
(b)
x 10
4
3.5
3
2.5
2
0
NAMW. Kg/Kmol
4.5
0.5
1.5
Time. hour
0.025
0.02
0.015
0.01
0.005
0
0
0.5
1.5
Time. hour
2.5
x 10
2.5
2
1.5
2.5
NAMW. Kg/Kmol
Fig. 3. Polymerization reactor control system response. (a) Servo control: (solid) RTD-A; (dashed) IMC-PID. (b) Regulatory control: (solid) RTD-A;
(dashed) IMC-PID.
0.5
1.5
Time. hour
2.5
0.5
1.5
Time. hour
2.5
0.04
0.03
0.02
0.01
0
0
Fig. 4. Polymerization reactor control system response: with noise. (a) Servo control: (solid) RTD-A; (dotted) IMC-PID. (b) Regulatory control: (solid)
RTD-A; (dotted) IMC-PID.
508
realistic process has also been demonstrated via simulation on a nonlinear polymerization reactor model. The
RTD-A controller was shown to provide better set-point
tracking and disturbance rejection performance than the
IMC-tuned PID controller, with and without measurement
noise.
Several issues remain outstanding, the most important of
which include theoretical robust stability analysisspecically how the choices of various parameter values jointly
and individually aect closed-loop stabilityand the development of systematic strategies for choosing the controller
parameters. A companion paper [29] presents a complete
discussion of these and related issues, following the preliminary results contained in [16].
References
[1] L. Desborough, R. Miller, Increasing customer value of industrial
control performance monitoringHoneywells experience, in: Proceedings of CPC VI Tucson, AZ, 2001.
[2] H. Koivo, J. Tanttu, Tuning of PID controllers: survey of SISO and
MIMO techniques, in: Proc. IFAC Intelligent Tuning and Adaptive
Control Symposium, Singapore, vol. 75, 1991, pp. 7580.
[3] P. Van Overschee, B. De Moor, The end of heuristic PID tuning, in
preprints of the IFAC Workshop on digital control: past, present and
future of PID control CBS: Terrassa, Spain, 2001, p. 687.
[4] D. Ender, Process control performance: not as good as you think,
Control Eng. (1993) 180.
[5] A.O. Dwyer, A summary of PI and PID tuning rules for processes
with time delay, Part I: PI controller tuning rules, in preprints of the
IFAC Workshop on digital control: past, present and future of PID
control, 175CBS: Terrassa, Spain, 2000.
[6] O. Lequin, M. Gevers, M. Mossberg, E. Bosmans, L. Triest, Iterative
feedback tuning of PID parameters: comparison with classical tuning
rules, Control Eng. Pract. 11 (2003) 10231033.
[7] M. Katebi, M. Moradi, Predictive PID controllers, IEE Proc.
Control Theory Appl. 148 (6) (2001).
[8] Q. Wang, C. Hang, X. Yang, Single-loop controller design via IMC
principles, Automatica 37 (2001) 20412048.
[9] I. Rusnak, Generalized PID controllers, in: The Proceedings of 7th
IEEE Mediterranean Conference on Control and Automation, 1999,
mED99, Haifa, Israel.
[10] M. Grimble, H1 PID controllers, Trans. IMC 13 (5) (1991) 112
120.
[11] K. Passino, S. Yurkovich, Fuzzy Control, Addison-Wesley-Longman, Menlo Park, CA, 1998.
strom, T. Hagglund, The future of PID control, Control Eng.
[12] K. A
Pract. (2001) 11631175.
[13] M. Morari, J. Lee, Model predictive control: past, present and future,
Comput. Chem. Eng. 23 (1999) 667682.
[14] S. Qin, T. Badgwell, A survey of industrial model predictive control
technology, Control Eng. Pract. 11 (2003) 733764.
[15] B. Ogunnaike, K. Mukati, Design, development and implementation
of an alternative structure for next generation regulatory controllers,
Annual AIChE Meeting, San Francisco, CA, 2003.
[16] K. Mukati, B. Ogunnaike, Stability analysis and tuning strategies for
a novel next generation regulatory controller, in: American Control
Conference, Boston, MA, 2004.
[17] K. Mukati, B. Ogunnaike, Tuning rules for the RTDA controller,
Annual AIChE Meeting, Austin, TX, 2004.
[18] G. Pannocchia, N. Laachi, J.B. Rawlings, A fast, easily tuned, SISO,
Model Predictive Controller, DYCOPS 7, Boston, MA, 2004.
[19] M. Soroush, K. Muske, Analytical model predictive control, Prog.
Syst. Control Theory 26 (2000) 163179.
509